The llm-coding-inference programme asks how many items a researcher must hand-code to get a valid interval from LLM-coded text. Its latest rule says: use the coder's own positive and negative cells as audit strata only if a 30% pilot shows the coder is informative (estimated J = sensitivity + specificity - 1 at least 0.3) and holds at least 20 human positives; otherwise pool.
The 0.3 was picked before looking. Milestone 6am sweeps it from 0.1 to 0.5 on cached Qwen2.5 zero-shot scores (GoEmotions admiration, neutral, fear, sadness; NaijaSenti negative share), 500 audits per cell.
- With the guard, admiration, fear and sadness do not change at all between 0.1 and 0.4.
- Neutral and NaijaSenti 1.5B, whose coders sit near J = 0.2, are the only cases where the cutoff changes which design runs, and half-widths differ by under 1%.
- 0.1 is too eager on a useless coder: NaijaSenti 3B at n = 400 is 3.9% wider than at 0.3.
- Coverage (Wilson on the effective size) is 92.6 to 96.6% in every cell, and the lowest cell is the same at every cutoff.
So 0.3 is not a knife edge; anything from 0.2 to 0.5 gives the same result here. What is still untested is an informative coder on a second corpus, which needs model weights this run could not download. Numbers: experiments/h6am_tau.txt.